CSR: Small: An Integrated Feedback Real-time Scheduling Framework for ECU and GPU-based Autonomous Driving Tasks

Ohio State University, The, Columbus OH

Investigators

Abstract

The rapid growth of autonomous driving systems (ADS) has resulted in two sets of new tasks for vehicle computing and control: 1) object detection and trajectory planning tasks that run machine learning algorithms mainly on on-board GPUs, 2) corresponding vehicle motion control tasks that run mainly on embedded Electronic Control Units (ECUs), such as path tracking. This project's novelty is an integrated ECU-GPU feedback real-time scheduling framework that helps ADS tasks meet their stringent deadlines, despite unexpected runtime execution time variations, while achieving 1) the maximum possible computing precision (and thus minimum tracking errors) for ECU-based motion control tasks, and 2) the highest possible recognition accuracy (and thus maximum safety) for GPU-based object detection and planning tasks. The project's broader significance and importance are its potential impacts on the designs of future autonomous vehicles, by substantially improving the timeliness of ECU and GPU-based ADS tasks, thus considerably enhancing the vehicle roadway safety and reducing the numbers of vehicle crashes, injuries, and fatalities. The new research challenges introduced by ADS cannot be properly handled by existing real-time scheduling solutions, because 1) ADS task execution times can vary significantly at runtime, 2) ADS heavily relies on GPUs that have more complex architectures than CPUs. To address those challenges, this project has several major research thrusts. First, it designs a two-tier ECU real-time scheduling algorithm that dynamically lowers the ADS execution times (with minimum precision degradation) within the allowed ranges for real-time guarantees. Second, it ensures the response times of GPU-based tasks are shorter than the desired deadlines, by adapting the GPU resources allocated to ADS tasks co-located on the same GPUs. Finally, this research jointly controls both ECU and GPU-based ADS tasks, as an integrated feedback real-time scheduling framework, to meet the end-to-end deadlines for all vehicle computing and control tasks. As autonomous vehicles are gradually becoming parts of our everyday lives, this timely project may produce findings that can offer effective approaches to substantially improving the real-time performance of ADS, thus enhancing the driving safety of future vehicles equipped with ADS. Dissemination and outreach are also planned to benefit industrial ADS systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.